Informatics Grand Rounds with Dr. Ahmed Hassoon

Why It Matters

Without coordinated, multi‑agent alignment, the expanding AI ecosystem could amplify clinical errors, inflate costs, and trigger systemic failures, jeopardizing both patient safety and health‑care sustainability.

Key Takeaways

• AI agents are proliferating across clinical, patient, insurance, and regulatory domains
• Uncoordinated AI interactions risk harmful equilibria like over‑testing
• Current alignment focuses on single‑agent behavior, not multi‑agent dynamics
• Proposed ecological framework aims to orchestrate agents toward system‑wide goals
• Robust, multi‑level alignment is essential to prevent catastrophic failures

Summary

The talk introduced a sweeping vision for AI alignment in medicine, emphasizing that today’s health‑care ecosystem is rapidly filling with autonomous agents—clinical decision‑support bots, patient‑facing assistants, insurance claim processors, and emerging regulatory AIs. Dr. Hassoon framed this proliferation as an "ecological framework" of interacting AI entities that could either reinforce existing inefficiencies or be orchestrated to overhaul the system.

He highlighted three core insights. First, each agent operates under its own incentive structure, creating hidden feedback loops that can drive undesirable outcomes, such as unnecessary imaging in an emergency‑department visit. Second, current alignment practices—pre‑training, supervised fine‑tuning, reinforcement learning with human feedback, and constitutional prompting—address only the behavior of individual models, not the collective dynamics when multiple agents act simultaneously. Third, real‑world analogues from finance (the 2010 high‑frequency‑trading flash crash), ride‑sharing (Uber price‑anarchy), and robotics (Okado warehouse fire) illustrate how unanticipated agent interactions can cause systemic failures.

Dr. Hassoon illustrated his points with a demo mapping LLM outputs onto a four‑dimensional space of safety, helpfulness, relevance, and fidelity, showing how surface‑level alignment can mask deeper misalignments. He quoted, "The question isn’t whether an agent is a lion; it’s what emerges when lions compete," underscoring the need to anticipate emergent equilibria. He also described his own research on testing alignment in clinical decision‑making and outlined a research agenda to develop multi‑agent orchestration protocols.

The implications are profound: health‑care leaders must shift from siloed AI deployments to system‑wide governance that aligns competing objectives, safeguards patient safety, and protects revenue cycles. Policymakers, hospital committees, and AI developers will need new frameworks—both technical and regulatory—to ensure that the collective behavior of AI agents advances, rather than undermines, the quality and cost‑effectiveness of care.

Original Description

This session with Dr. Ahmed Hassoon will introduce the topic of AI Alignment, what it means, and how it is currently done. #aialignment #johnshopkins #bioinformatics #healthinformatics https://bids.jhmi.edu/about/events-also-past-grand-rounds/

This session covers:

-Introducing the topic of AI Alignment, what it means and how it is currently done

-Explaining the limitation of current alignment approaches

-Describing what alignment means in medical practice and if it is achievable

-Demonstrating how the practice of medicine will change in the future

-Providing examples of what kind of alignment we should start thinking about

-Identifying other interesting things AI can do in medicine such as an overview of AI in discovery